Refactor segmentation evaluation logic into a reusable module

## Problem

Currently, evaluation logic appears to be implemented within individual segmentation model classes (Torch, TensorFlow, ONNX), which may lead to duplicated logic and inconsistencies across backends.

## Proposed Solution

Introduce a reusable evaluation module that:

- Handles dataset iteration
- Calls model.predict() for inference
- Computes metrics in a centralized manner

This would allow model classes to focus purely on inference while keeping evaluation logic consistent across frameworks.

## Scope

This proposal focuses on extracting common evaluation patterns without breaking existing functionality.

## Expected Impact

- Reduced code duplication
- Improved maintainability
- Consistent evaluation behavior across different model backends